ComfyUI Extension: ComfyArt

ComfyArt Introduction

ComfyArt is a free, open-source AI design workspace built into ComfyUI. It brings an infinite canvas, editable design tools, and AI generation together so you can develop a project without constantly switching between separate applications.

Think of it as a digital studio table: you can place reference images, create artwork, arrange text and shapes, try different layouts, and prepare your finished design for export. Alongside image creation, ComfyArt supports video, music, voiceovers, sound effects, and textured 3D models.

You can use ComfyArt manually or work with an optional AI design Agent. The Agent can use tools to create and edit content on your canvas, rather than only suggesting what you should do.

ComfyArt helps you:

• Keep your creative work together. Organize references, generated media, editable layouts, and conversations within projects.
• Move beyond a single generated image. Add precise typography, arrange elements, remove backgrounds, and turn artwork into a finished composition.
• Revise existing designs. Use image-editing tools or reconstruct a flat poster as editable elements.
• Reuse your visual identity. Keep brand materials, characters, products, and references available for future work.
• Choose where generation happens. Run supported workflows on your own computer, another ComfyUI computer on your local network, or Comfy Cloud. Local generation does not require a ComfyArt account, credits, or cloud processing. Optional Comfy Cloud generation requires a paid Comfy plan and an API key. The Agent’s processing location and account requirements depend on the provider you choose; selecting a hosted Agent does not make its requests local.

ComfyArt canvas and workflow preview

Watch the 38-second ComfyArt overview

An example of the Agent building an Instagram post on the ComfyArt canvas

For illustrated documentation covering every feature, visit comfyart.org (https://comfyart.org).

How ComfyArt Works

The canvas is your workspace

ComfyArt uses an infinite canvas built with the tldraw SDK and React. You do not need to understand those technologies to use it: the important idea is that you have room to arrange ideas and finished work side by side.

Within a project, you can use:

• Pages to separate parts of your work.
• Artboards to define individual compositions, such as a social post or presentation slide.
• Layers to organize the objects within a design.
• Media cards to work with generated content such as images and playable audio. An artboard is like a sheet of paper on a much larger desk. You can keep several designs nearby while giving each one its own dimensions and layout.

AI tools run prepared ComfyUI workflows

A workflow is a prepared sequence of processing steps. For example, removing a background involves loading the right model, analyzing your image, and producing a cut-out result.

ComfyArt’s AI generation and processing tools use prepared workflows in the workflows/ folder. You work through the canvas controls rather than manually connecting each workflow’s nodes.

Different tools need different models. Having an image generator available does not automatically mean that video, background removal, or 3D generation is also ready.

References guide the result

A written prompt describes what you want. A reference image can provide additional visual guidance, while a structure mode can turn that reference into a simplified guide.

For example:

| Guide | What it represents | Example use | |---|---|---| | Line art or edges | Outlines and boundaries | Guide a new illustration with an existing composition | | Depth | Which areas appear nearer or farther away | Guide the spatial arrangement of a scene | | Pose | The arrangement of a body | Guide a character’s posture | | Surface normals | Which direction surfaces face | Provide information about an object’s form | | Segmentation | Regions representing different parts of a scene | Guide the placement of major visual areas |

These guides are like tracing aids: they provide structure rather than acting as finished artwork. Available reference modes depend on the workflow and its supporting models.

ComfyArt uses ControlNet Auxiliary Preprocessors for reference processing. These preprocessors create guidance images; they are not image generators by themselves.

Choose where the AI work runs

| Processing location | How it works | When it is useful | |---|---|---| | This computer | ComfyUI runs workflows using your local models and hardware. | You want local processing and have the required resources. | | A local-network computer | ComfyArt sends a job to another configured ComfyUI machine and brings the result back. | You have access to another workstation or want to distribute work. | | Comfy Cloud | Inputs are uploaded for processing, and results are downloaded into your ComfyUI output folder. | You want supported generation without relying on a local GPU or downloading those models locally. |

Choose the default destination under Settings → Generate on.

For configured local-network computers, Auto selects the least busy compatible machine, including your own computer. It skips computers that are offline or lack required nodes or model files. Individual generator cards can override the default through Run on.

The optional Agent works through tools

The Agent acts more like a design assistant with access to your workspace than a plain chatbot. It can inspect the canvas and use tools to generate media, edit images, arrange text and shapes, create mockups, and work with other supported features.

For example, you might ask:

“Create a product announcement using this bottle image, a cream background, and a bold headline. Keep space for a short description.”

This is an example request, not a guaranteed result. What the Agent can complete depends on its capabilities and the tools and models available.

The Agent communicates through MCP, a standard connection that lets an AI assistant use application tools. MCP and the Agent are off by default and can be enabled in Agent Settings.

Keep the project’s ComfyArt tab open while the Agent works. Canvas actions are carried out by that open tab.

ComfyArt Features

Projects, preferences, and saved work

You can start a design from a sentence or a skill, then browse and manage your projects from the home screen.

Workspace preferences include:

• English or Chinese interface language.
• Theme selection.
• Chat-panel side. The interface initially follows ComfyUI’s language.

Projects, snapshot history, and Agent chats are saved in a SQLite database at:

text user/__comfyart/data.db

Older projects stored as JSON files under comfyart/projects/ are imported automatically the first time you open ComfyArt. Their original files remain in place.

Canvas layout and organization

The canvas includes:

• Pages and artboards with size presets.
• Fill, stroke, and corner-radius controls.
• Auto layout.
• A layers panel with dragging, renaming, and range selection.
• Snapshot history.
• Comments that can become edits or be sent to the Agent with the marked area. For example, you can keep a product image, headline, and supporting text as separate objects. This makes rearranging a composition easier than repeatedly generating an entirely new poster.

Comments let you connect feedback to a specific location. A note such as “replace this background” can be associated with the relevant area instead of describing its position in chat.

Rich text and vector drawing

You can combine different fonts, sizes, colors, and emphasis within one text box. Supported formatting includes bold, italic, and underline.

The font picker includes brand, common, and local fonts.

Vector tools include:

• Solid, linear-gradient, radial-gradient, and image fills.
• Transparent color stops and blend modes.
• A pen tool with node editing.
• Boolean shape operations: union, subtract, intersect, and exclude. For example, a transparent gradient can fade a shape over a photograph, while subtracting one shape from another can create a cut-out graphic.

Image generation

Image generators support dimensions measured in real pixels, model selection, references, structure modes, color palettes, and base images.

These controls serve different purposes:

• Dimensions define the generated image’s size and proportions.
• Model selection chooses the generation workflow.
• References provide visual guidance.
• Structure modes provide guidance such as outlines, depth, or pose.
• Color palettes communicate a color direction.
• Base images provide existing material to work from. For a coordinated campaign, you might use the same palette and product references across multiple compositions. Results still depend on the chosen model and workflow.

Image editing and finishing

| Tool | What you can use it for | |---|---| | Quick edit | Request changes to an existing image. | | Upscale | Increase image resolution with the 4× UltraSharp workflow. | | Remove background | Isolate a subject for use in another composition. | | Eraser | Remove unwanted content and fill the affected area using LaMa. | | Crop | Reframe an image, including with platform presets. | | Flip and rotate | Change orientation. | | Non-destructive adjust | Adjust appearance without permanently replacing the source through that adjustment. | | Inpaint | Regenerate a selected area within an image. | | Move object | Reposition an object within an image. | | Multi-angle | Generate another camera view with the supported Qwen workflow. | | Expand | Generate content beyond the existing image boundaries. | | Split layers | Separate a flat image into layers with Qwen Image Layered. | | Edit text | Modify text contained within an image. | | Style transfer | Apply a different visual style. | | Vectorize | Convert image material into vector artwork. |

Inpainting works inside an image; expanding, also called outpainting, works beyond its current edges. For example, inpainting could replace a small object, while expanding could adapt a tight composition to a wider format.

Edit Elements: make a flat poster editable

A flat poster is usually one image, even if it visually contains text, a product, a logo, and a background.

Edit Elements reconstructs these parts into an editable artboard using:

• OCR—text recognition—to identify text layers.
• Segmentation to isolate visual elements such as products and logos.
• Background cleanup to fill areas behind removed elements. You can then edit the reconstructed design and re-render it.

This combines RapidOCR, SAM3, and LaMa. It is different from Split layers, which uses Qwen Image Layered. Reconstructing a flattened image is not the same as recovering its original design file, so review the extracted content before finalizing it.

Product mockups

Mockups place a design onto a product photograph. The design follows folds and shading in real time, helping you evaluate how it might look on the photographed surface.

You can then bake the placement into an image or blend it with AI.

For example, you could preview a graphic on a photographed fabric item before producing a finished presentation image. The workflow uses depth estimation and background removal to understand the photograph.

Video creation and editing

ComfyArt supports video generation from:

• Text.
• First and last frames.
• Up to nine reference images.
• A storyboard.
• Reference media, including images, video, and audio. Documented generation options include 5–15 seconds at 768p or 1080p. Other supported workflows include animation, video extension, motion control, and lip sync.

The video editor combines several clips and music for export using ComfyArt’s Timeline Render node.

For example, you can generate short shots, assemble them, and add music rather than trying to produce a complete sequence in one generation.

Audio and 3D

Audio tools generate music, voiceovers, cloned voices, and sound effects. Results appear as playable cards.

Image-to-3D tools create textured 3D models from a source image. Pixal3D also has a multi-view workflow for working from multiple views.

Use voice cloning only with appropriate permission and rights to the reference voice.

Assets and brand materials

Asset organization includes:

• My Assets.
• A full-page brand kit editor.
• Characters.
• Products.
• A reference board. These help you keep reusable material close to your work. For example, you can organize product references and brand materials while exploring several promotional layouts.

Export and navigation

Supported design export formats are:

• PNG
• JPG
• PDF
• Editable PPTX PNG and JPG provide image outputs; PDF is useful for sharing document-style layouts; editable PPTX supports further presentation editing.

Open ComfyArt from ComfyUI’s top bar or, with the default local address, at:

text http://127.0.0.1:8188/comfyart/

Use Back to ComfyUI in the home rail or the project logo menu to return.

ComfyArt Models

ComfyArt uses different models for different tasks. The tables below list the model families and roles documented in the supplied workflows.

No comparative speed, memory, or quality benchmarks are provided. Treat alternative generators as options to evaluate with your own prompts, rather than assuming one is always best.

Image generation models

| Model | Documented role | When to choose it | |---|---|---| | Qwen Image 2.1 | Image generation, quick edits, text editing, expansion, comment edits, and multi-angle generation with an additional camera-angle LoRA. | You want the documented Qwen generation-and-editing workflow. | | Z-Image Turbo | Alternative text-to-image generation. | You want to try another generator for your prompt. | | FLUX.2 Klein 9B | Alternative text-to-image generation. | You want to compare the FLUX.2 workflow with other image generators. | | FLUX.3 Klein | Alternative text-to-image generation. | You want to evaluate the documented FLUX.3 workflow. | | MOSS Image | Alternative text-to-image generation. | You want another generation option to compare. | | Qwen Image 2512 | Image generation, with additional ControlNet requirements for structure guidance. | You want this Qwen workflow or its supported reference modes. | | Ideogram 4 | Image generation using a structured caption written by the default Agent. | You want the Ideogram workflow and have the required Agent available. |

Although the original feature summary mentions five local image models, its detailed workflow list names the seven options above.

Important dependencies:

• Qwen Image 2512 uses the ComfyArt Precise UNET Loader locally to avoid a documented FP8-loading problem.
• Qwen Image 2512 structure guidance also needs Qwen-Image-2512-Fun-Controlnet-Union-2602.safetensors.
• Multi-angle generation uses the Qwen Image 2.1 camera-angle LoRA.

Image-processing models

| Model | Role | |---|---| | 4× UltraSharp | Image upscaling. | | BiRefNet | Background removal; also used by mockup and 3D workflows. | | LaMa | Erasing and background cleanup, including within Edit Elements. | | Qwen Image Layered | Splitting a flat image into layers. | | SAM3 | Isolating elements for Edit Elements. | | Depth Anything V2 | Estimating depth for mockup placement. |

RapidOCR supplies local text recognition for Edit Elements. Supporting reference preprocessors generate guides such as line art, depth, pose, normals, and segmentation; not every preprocessor or mode is available in every workflow.

Video models

MiniMax H3 powers the documented video workflows:

• Text-to-video and image-to-video for generating clips, animation, and extension.
• Reference-to-video for reference images, video, and audio.
• Motion control, with an additional control model and pose-detection models.
• Lip sync, using the image-to-video workflow with a guide. Motion control additionally uses the MiniMax H3 control patch, SDPose Wholebody, and RT-DETR models.

Audio models

| Model | Role | |---|---| | MiniMax Music 3 | Music generation. | | ACE-Step 1.5 | Alternative music generation. | | YuE2 | Alternative music generation. | | Stable Audio Open 1.0 | Sound-effect generation. | | Kokoro | Local text-to-speech voiceovers. | | IndexTTS2 | Local voice cloning through ComfyUI Easy IndexTTS2. | | Qwen3-TTS | Voiceovers and voice cloning on Comfy Cloud instead of Kokoro and IndexTTS2. | | Gemma 4 | Transcription for voice input. |

The underlying Kokoro nodes support speaker selection, speech speed, language, and speaker blending. Their availability in the ComfyArt interface depends on the exposed workflow controls. See the Kokoro voice list for supported voices.

3D models

| Model | Input and role | |---|---| | TRELLIS.2 | Creates a textured 3D model from an image. | | Pixal3D | An alternative image-to-3D workflow using additional geometry estimation. | | Pixal3D Multiview | Creates a 3D model from multiple views. |

The Pixal3D workflows share supporting components with TRELLIS.2. Use the multi-view workflow when you have multiple views to supply; it requires its own multiview model.

Agent options

Agent models are separate from the models that generate your media.

| Agent option | Requirement | |---|---| | Claude Code | Available and logged-in claude command. | | Codex | Available and logged-in codex command. | | Qwen Code | Available qwen command with a configured model. | | DeepSeek Harness | Available dsh command, signed in through its application. | | Local LLM | An OpenAI-compatible endpoint supporting tool calling, configured in Agent Settings. |

You can choose the Agent per chat. ComfyArt does not store credentials.

Some helper tasks—including storyboard splitting, OCR-related calls, brand and memory extraction, reference notes, and Ideogram caption preparation—make individual calls to the default Agent. These are distinct from the local RapidOCR text-recognition service.

Troubleshooting ComfyArt

The Agent does not respond or cannot change the canvas

1. Open Agent Settings.
2. Confirm that both MCP and the Agent are enabled.
3. Check that the selected provider is available and signed in, or that your Local LLM endpoint is configured and supports tool calling.
4. Keep the correct project’s ComfyArt tab open.
5. If your Local LLM requires authentication, confirm that its API key was made available before ComfyUI started. Closing the chat panel does not stop an ongoing chat. Closing the project tab prevents the browser-side canvas tools from carrying out actions.

An image or video workflow reports missing models or nodes

1. Identify the selected tool and model.
2. Check whether the job is running locally, on a network computer, or on Cloud.
3. Check that the destination supports all components required by that workflow.
4. For a network computer, check model file names, not just folder locations: ComfyArt matches files by name.
5. If the destination is unavailable, choose another compatible destination. Different tools have separate requirements. For example, successful image generation does not establish that LaMa erasing or MiniMax H3 video generation is available.

Qwen Image 2512 produces noise instead of an image

The author documents a conflict between unscaled FP8 checkpoints and ComfyUI’s --fast fp8_matrix_mult loading path.

1. Confirm that the problem affects Qwen Image 2512.
2. For local or local-network processing, use its supplied workflow with the ComfyArt Precise UNET Loader.
3. If using another network computer, confirm that ComfyArt is available there so its loader can run. The loader bypasses the problematic path for the model it loads. On Comfy Cloud, this workflow uses the core UNETLoader instead.

A reference mode or mockup takes longer on its first run

Some supporting weights download automatically when first needed.

1. Allow time for the initial download.
2. Check internet access if the process cannot complete.
3. Check ComfyUI’s processing log for a failed download or missing component. Documented first-use downloads include MLSD, HED, normal and segmentation annotators, and Depth Anything V2 for mockups. The underlying Kokoro integration also downloads its model and speaker metadata on first use.

Edit Elements cannot reconstruct a poster

1. Confirm that text recognition is available through RapidOCR.
2. Confirm that the SAM3 model is available for element isolation.
3. Confirm that LaMa and its supporting inpaint nodes are available for cleanup.
4. Review recognized text and extracted objects before re-rendering. A failure in one stage can affect the overall reconstruction. Edit Elements does not recover the poster’s original source file.

Comfy Cloud generation fails

1. Confirm that your Cloud API key belongs to a paid Comfy plan.
2. Confirm that COMFY_API_KEY was available before ComfyUI started.
3. Select Settings → Generate on → Comfy Cloud.
4. Remember that submitted images and audio must be uploaded for Cloud processing. Manage keys at Comfy’s API-key page. Cloud results are downloaded into output/comfyart/ with a cloud-<job> suffix.

A local-network computer is not receiving jobs

1. Check its address and port under Settings → LAN instances….
2. Confirm that it is online and accessible. The remote ComfyUI must accept network connections.
3. Check that it has the required nodes and model file names.
4. Check Settings → Generate on and the card’s Run on override.
5. If using Auto, remember that ComfyArt may choose a different compatible, less busy computer. Network results are downloaded into output/comfyart/ with a <machine>-<job> suffix.

Frequently asked questions

Do I need the Agent to generate images or videos? No. Image and video modes work without the Agent. However, certain helper operations, including Ideogram’s structured caption preparation, use the default Agent.

Is everything processed on my computer? Only when you choose local processing and local services. Cloud jobs upload inputs, network jobs send files to the selected computer, and Agent requests follow your chosen provider.

Can I use ComfyArt without a local GPU? Comfy Cloud is a documented alternative for running workflows. It requires a paid Comfy plan and API key.

Where are my projects and conversations saved? Projects, snapshot history, and Agent chats are saved in user/__comfyart/data.db.

Can I recover older JSON projects? ComfyArt automatically imports projects from comfyart/projects/ on first opening and leaves the original files in place.

Why is there a “Get a license for production” badge? The canvas uses the separately licensed tldraw SDK. The author documents that it runs in development mode without a license key on localhost or plain HTTP and displays this badge.

Can an external AI assistant use ComfyArt? Yes. Enable MCP and connect a compatible client to the project-specific endpoint. The MCP guide (https://comfyart.org/guides/mcp) covers supported clients and tool behavior.

Learn More about ComfyArt

Start with visual documentation

• ComfyArt documentation (https://comfyart.org) — The primary resource, with screenshots for every feature.
• 38-second overview — A quick visual introduction to the workspace.
• ComfyArt repository — Source documentation and supplied workflows.

Explore AI-assisted design

• ComfyArt MCP guide (https://comfyart.org/guides/mcp) — Connect external assistants, understand tool timeouts, and explore the full tool list. Includes guidance for Claude Code, Codex, Gemini CLI, Qwen Code, Cursor, VS Code, and Claude Desktop.
• ComfyTV — A related media-workbench project from which the author adapted ComfyArt’s Agent backend. Its feature list is separate from ComfyArt’s.

Understand supporting creative tools

These resources describe the underlying integrations. Their full feature sets are not necessarily exposed in ComfyArt.

• ControlNet Auxiliary Preprocessors — Examples and explanations of structural guides such as line art, depth, and pose.
• ComfyUI Inpaint Nodes — Background on LaMa erasing and masked-image processing.
• ComfyUI Kokoro — Local speech-generation controls and examples.
• ComfyUI Easy IndexTTS2 — Reference material for the local voice-cloning integration.
• Comfy Cloud — The optional hosted workflow-processing service.

ComfyArt License

ComfyArt is MIT licensed. The author adapted its Agent backend from ComfyTV’s bot, as described in the project’s licensing notes.

The tldraw SDK has a separate license, included in licenses/tldraw-LICENSE.md. ComfyArt’s MIT license does not replace that SDK’s terms.

Supporting models and integrations may also have their own licenses. Review the relevant terms when choosing tools for your work, particularly for commercial projects or publishing generated content.